# PUSHING GUIDE — how to formally run a Pushing enquiry (the Logicvid method) *2026-08-12. The comprehensive, reusable guide to the Logicvid / Pushing method, distilled from the source files in `pushing/_source/` and the 30+ worked sessions in `research-library/recognition/`. It answers: how do you mechanically, rigorously, FORMALLY push a text to its bottom — and how do you turn the resulting logical arguments into auditable truth-packets you can cite in essays?* > **For an agent told "create a pushing file for ":** read > `AUTONOMOUS_PUSHING_AGENT_SPEC.md` — it is the self-contained instruction (engineer the > question-asking context first, run the loop to natural endpoints, pivot until repeats, output > truth-packets, do NOT commit to git). This guide is the *method*; the agent spec is the *operator*. --- ## PART 1 — WHAT PUSHING IS ### 1.1 The one rule > **Hound the text with "why," and force its OWN reasoning out. Our frameworks only supply the > questioning; the answers must come from the text.** The extracted passages are the source of truth. Each question stores the full quoted passage (more than needed) so the text's own argument can be re-read and re-pushed. Frameworks (aperture, MEPIT, pure-thesis, Solms/valence) enter ONLY after the text has spoken. ### 1.2 The method is a graph-growth machine From `logicvidsmethod.md`, the two loops that compose: ``` SANSKRIT → claims → definitions → dependencies → proof-or-boundary → GRAPH (decomposition) graph tension → paradox → hidden premises → branches → research → NEW GRAPH (question-growth) ``` Pushing is the **question-growth** loop on top of the **decomposition** loop. It is not separate from the formal engine — it is the layer that grows the graph the formal engine then proves over. ### 1.3 The three reliability levels (from logicvidsmethod) | Level | Feasibility | What it yields | |---|---|---| | **1. Structural map** | absolutely feasible | per-āhnika/chunk: problem, primitives, presuppositions, new claim, inferential work, dependencies, tensions | | **2. Hard metaphysical backbone** | feasible, slower | a minimal claim-chain, each arrow with passages + commentary + hidden premises + objections + confidence | | **3. Definitive reconstruction** | not from one translation alone | needs the commentarial corpus (Jayaratha, Tantrasāra, Utpaladeva, Pratyabhijñā) | **Advice:** build Level 1–2 first. Do not attempt Level 3 until the corpus (commentaries + parallels) is assembled. --- ## PART 2 — THE FORMAL PROCEDURE ### 2.0 Set up the enquiry (the repeatable artifact) Each enquiry is one folder (from logicvidsmethod): ```text inquiries// root.json the question + source scope + core tension + hidden premises + status source-spans.json the quoted passages (the text speaks) reconstruction.md Pass A: the strongest coherent reconstruction prosecution.md Pass B: destroy it provenance.md Pass C: every claim → a passage branches.md the residual questions logicvid.md the compiled enquiry (the output) ``` `root.json`: ```json { "question": "If I must already be Śiva to recognize myself as Śiva, how can I recognize myself as not-Śiva?", "source_scope": ["Tantrāloka", "Tantrasāra", "Īśvarapratyabhijñā"], "core_tension": ["identity is invariant", "self-knowledge is presently limited"], "hidden_premises": [], "status": "seed" } ``` ### 2.1 The three passes (the anti-cheat) **Pass A — Construct.** Produce the strongest coherent reconstruction from the passages. Output: primitives · claims · dependency chain · hidden premises · chapter function · tensions · questions. **Pass B — Destroy.** A separate agent is FORBIDDEN from improving it; its only job is to find: unsupported entailments · conflated levels · translation dependence · missing intermediate claims · passages that resist · rival readings · false formalization · contradictions hidden by vague wording. The construct agent repairs only what survives. **Pass C — Provenance audit.** Every explicit claim has a passage · every derived claim lists premises · every cross-source claim preserves direction · every contradiction survives scope separation · every Sanskrit term keeps tradition-local meaning. ### 2.2 The round-structure of a session (from the PUSHING transcripts) Each session is a chain of rounds — question → the text's answer → the deeper question it forces: ``` ROUND N The question (the sharp "why") The text's answer (restated EXACTLY as it argues — no strawman) The new deeper question it forces > PENETRATION N: (the exact spot where the text asserts but does not prove) ``` The example (Q1-reflexivity): "If the perceiver is Śiva-universal, why is blue manifest to me and not everyone?" → the text replies via Sāṁkhya-refutation → the deeper question: "why is the empirical 'me' a contraction of the Light rather than a recipient of it?" → **PENETRATION 1:** the text asserts the collapse but does not prove it (the quantifier problem). --- ## PART 3 — THE HONESTY RULES (the discipline that makes it trustworthy) ### 3.1 The licensed-vs-not rule (from logicdog) Always split what the text **licenses** from what it **does not**. Example ("Is a dog Śiva?"): - **Ontological recognition** — the text licenses this: a dog enacts aham-vimarśa (contracted first-person), so "the dog manifests Śiva's recognitive structure." - **Explicit liberating recognition** — the text does NOT license this: no reliable passage says a dog attains Pratyabhijñā liberation as a dog. State it as unsupported. ### 3.2 The relabelling accusation (from logicdog §5) The naturalist's cleanest objection: `vimarśa` is "a metaphysical relabelling of multiscale self-organization" — poetic description, no predictive mechanism. **You must face this head-on**: Abhinavagupta can *accept every empirical mechanism* and still hold that it does not explain why manifestation is self-apprehending. The reply is not "add mechanism" but "the mechanism presupposes the manifest." ### 3.3 Three errors to refuse (from logicvidsmethod) - **Compression error** — forcing distinct claims into one clean system. - **Attribution error** — giving Abhinavagupta a later commentator's repair. - **Bridge error** — treating similar terms as identical (prakāśa = phenomenal consciousness) when they are *probes, not identities*. ### 3.4 Never collapse levels (from logicframework) Use the 6-level frame and never collapse: ``` M → D → B → N → W → R manifestation → differentiation → bounded embodiment → normative agency → world-model → recognition ``` Science explains D→B→N→W; Pratyabhijñā proposes M→D and interprets R as W→awareness-of-M. The dispute is whether M is generated by, identical with, or prior to the later levels. **Never write "consciousness = integration."** --- ## PART 4 — TURNING ARGUMENTS INTO AUDITABLE TRUTH-PACKETS This is the key idea: **treat a logical argument almost like a translation** — a derived object with an auditable path, resolvable to source, graded by how well it is proved. ### 4.1 The argument as a first-class, translation-like object Just as a translation has source_spans → target_spans → decisions → evidence, a logical argument has: ``` pt:argument:: { work_id title kind "reductio" | "analogy" | "identity" | "entailment" | "decomposition" premises [ { text, passage_ids } ] inference the typed move conclusion { text, passage_ids } tension_id the PUSHING question it resolves provenance resolved passages (auditable path) proof optional — the truth-engine trace (Lean / Nyāya) status MACHINE_DRAFT → REVIEWED | PROVED | OUTSIDE_FORMAL | HOLLOW } ``` The **auditable path** mirrors a translation's provenance: conclusion → inference → premises → each premise resolves to its passage (via `/api/resolve`). You can trace any claim back to the Sanskrit. ### 4.2 Strength-graded claims (the "different strength" idea) Not all arguments prove equally. Attach a **claim strength** so essays can cite at the right level: ``` PROVED a formal proof exists (truth engine / Lean) REVIEWED human-review has accepted the reconstruction + provenance WELL_SUPPORTED premises resolve, inference is sound, no surviving prosecution PLAUSIBLE a coherent reconstruction with a live objection (the tension stands) SPECULATIVE a probe — explicitly NOT asserted as the text's claim ``` A claim's strength is **derived from the argument object's status + the surviving objections**, not hand-waved. An essay then says "the text's position (WELL_SUPPORTED, prem. A, B, C)" vs "a possible reading (SPECULATIVE)" — so the reader knows exactly how load-bearing each claim is. ### 4.3 Why this is like a translation | Translation | Argument | |---|---| | source_spans | premises (each resolves to a passage) | | target_spans | the claim / conclusion | | decisions | the inference + hidden-premise choices | | evidence | the quoted passages | | review_state | claim strength (PROVED → SPECULATIVE) | | resolve kernel | trace any claim back to Sanskrit | This makes the argument a **truth-packet**: a self-contained, auditable, strength-graded object you can cite in an essay (SHOW EVIDENCE → the argument → its premises → the passages), exactly like citing a translation. --- ## PART 5 — THE COMPOUNDING PIPELINE (recap) ``` PUSHING enquiry (finds a tension, quotes the passages) → resolve passages (/api/resolve + published store) → FORMAL LOGICAL ARGUMENT (the truth-packet: premises/inference/conclusion + auditable path) → TRUTH ENGINE (nyāya/Lean: PROVED / OUTSIDE_FORMAL / HOLLOW) → ESSAY (cites the argument at its correct claim-strength) → LEARNING (from the essay) → back to PUSHING the next tension ``` Everything is tracked on the source hub (`/api/hub`) — `pt:hub:::` with `passage_ids`. Nothing is orphaned. --- ## PART 6 — ADVICE (what actually works) 1. **Start at Level 1–2**, not Level 3. Build the structural map + a backbone of a few genuine cruxes before attempting the definitive reconstruction. 2. **One enquiry = one folder.** The repeatable artifact shape (§2.0) is the whole method made tangible. Get 5–10 excellent enquiries before exposing them via MCP/API. 3. **Face the relabelling accusation on every tension.** It is the strongest objection; meeting it is what separates a real penetration from a gloss. 4. **Store more passage than you need.** The text must be re-readable and re-pushable. 5. **Pass B must be a separate agent that cannot improve** — its only job is to destroy. This is the anti-cheat that makes the reconstruction trustworthy. 6. **Let the strength be derived, not claimed.** PROVED / REVIEWED / WELL_SUPPORTED / PLAUSIBLE / SPECULATIVE comes from the proof status + surviving objections, so no essay overclaims. 7. **Never collapse levels or terms.** The 6-level frame and the term-probes-vs-identities rule keep the analysis honest. --- ## PART 7 — THE SEED (what's on disk) - **Source files:** `pushing/_source/` (logicvidsmethod, logicframework, logicframework2, logicdog, logic5/6/7, logicvid3, postmortem, PUSHING-TANTRALOKA, PUSHING-IPVV). - **Worked sessions:** `recognition/pushing-tantraloka/LOGICVID-session-*.md` (30+) + `recognition/pushing-ipvv/LOGICVID-session-*.md`. - **The hub:** `data/corpus/hub.ts` already tracks `pt:hub:ipvv:pushing:main` and `pt:hub:tantraloka:pushing:main`. This guide is the how-to; the sessions are the evidence it works; the argument-object schema is the natural next build (Agent 1).